# Load the saved meta graph and restore variables
saver = tf.train.import_meta_graph("{}.meta".format(checkpoint_file))
saver.restore(sess, checkpoint_file)
# Get the placeholders from the graph by name
inputs = graph.get_operation_by_name("inputs").outputs[0]
labels = graph.get_operation_by_name("labels").outputs[0]
vocab_embedding = graph.get_operation_by_name("vocab_embedding").outputs[0]
discard_last_prediction = graph.get_operation_by_name("discard_last_prediction").outputs[0]
# Tensors we want to evaluate
predictions = graph.get_operation_by_name("accuracy/predictions").outputs[0]
perplexity = graph.get_operation_by_name("perplexity/perplexity").outputs[0]
# Generate batches for one epoch
batches = preprocess_helper.batch_iter(list(zip(x_test, y_test)), FLAGS.batch_size, 1, shuffle=False)
# Construct the embedding matrix
vocab_emb = np.zeros(shape=(FLAGS.vocabulary_size, FLAGS.embedding_dimension))
w2v_model = gensim.models.KeyedVectors.load_word2vec_format(FLAGS.path_to_word2vec, binary=False)
for tok, idx in vocab.items():
if FLAGS.use_word2vec_emb and tok in w2v_model.vocab:
vocab_emb[idx] = w2v_model[tok]
else:
vocab_emb[idx] = generated_embeddings[tok]
# Collect the predictions here
all_perplexity = []
if(FLAGS.verbose_for_experiments):
print("verbose_for_experiments: Only the perplexity will be shown for each sentence")
val_sample_index = -1 * int(FLAGS.val_sample_percentage * float(len(labels)))
x_train, x_val = data[:val_sample_index], data[val_sample_index:]
y_train, y_val = labels[:val_sample_index], labels[val_sample_index:]
# Summary of the loaded data
print('Loaded: ', len(x_train), ' samples for training')
print('Loaded: ', len(x_val), ' samples for validation')
print('Training input has shape: ', np.shape(x_train))
print('Validation input has shape: ', np.shape(x_val))
print('Training labels has shape: ', np.shape(y_train))
print('Validation labels has shape: ', np.shape(y_val))
# Generate training batches
batches = preprocess_helper.batch_iter(list(zip(x_train, y_train)),
FLAGS.batch_size, FLAGS.num_epochs)
print("Loading and preprocessing done \n")
# Define the model and start training
with tf.Graph().as_default():
session_conf = tf.ConfigProto(
allow_soft_placement=FLAGS.allow_soft_placement,
log_device_placement=FLAGS.log_device_placement,
inter_op_parallelism_threads=FLAGS.inter_op_parallelism_threads,
intra_op_parallelism_threads=FLAGS.intra_op_parallelism_threads)
sess = tf.Session(config=session_conf)
with sess.as_default():
# Initialize model
lstm_language_model = LSTMLanguageModel(
vocabulary_size=FLAGS.vocabulary_size,